CN115861804A - SAR image countercheck sample detection system and method based on optimal feature attribution selection - Google Patents
SAR image countercheck sample detection system and method based on optimal feature attribution selection Download PDFInfo
- Publication number
- CN115861804A CN115861804A CN202211504977.8A CN202211504977A CN115861804A CN 115861804 A CN115861804 A CN 115861804A CN 202211504977 A CN202211504977 A CN 202211504977A CN 115861804 A CN115861804 A CN 115861804A
- Authority
- CN
- China
- Prior art keywords
- individual
- sample
- image
- individuals
- population
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a synthetic aperture radar image confrontation sample detection system and method based on multi-target optimal feature attribution selection. The method comprises the steps of collecting historical SAR images from an SAR system monitoring database, using the historical SAR images as an input data set after data normalization and normalization, using the number of sub-samples generated in the characteristic analysis process based on sliding scanning and the area under a working characteristic curve of a subject of a logistic regression model as optimization targets, designing a characteristic scanning block parameter optimization platform based on a multi-target optimization method, and obtaining an optimal characteristic attribution scanning block and an optimal regression model. And performing countermeasure sample online detection on the real-time SAR image data in the SAR system real-time database by using the optimal regression model. The technology of the invention can automatically obtain the optimal feature analysis granularity according to different scenes, efficiently realize the detection of various confrontation samples in the SAR image recognition field, and improve the calculation efficiency and AUC performance index of SAR image confrontation detection.
Description
Technical Field
The invention relates to a confrontation sample detection technology in the field of synthetic aperture radar image identification safety, in particular to a system and a method for detecting a confrontation sample based on multi-target optimal characteristic attribution selection.
Background
Synthetic Aperture Radar (SAR) as an active remote sensing system can acquire a high-resolution image of a target area at any time without being influenced by weather factors such as cloud and fog, illumination and the like. Compared with traditional passive sensors which are easily influenced by weather, such as optics, infrared sensors and the like, the SAR provides an effective solution for a target detection task, so that the SAR is widely applied to the fields of mountain monitoring, maritime management and the like. In recent years, the development of deep learning technology further improves the precision of the SAR identification system, promotes the deployment and popularization of the SAR system, and meanwhile, the SAR system faces the threat of the deep learning safety problem. In many security problems, countersample attacks which cause model misclassification seriously affect the robustness of the SAR recognition model, and the recognition effect of the SAR recognition model is reduced. In some important mission scenarios, such as military monitoring, one false identification may be sufficient to have unacceptable consequences. This makes modern SAR systems require a corresponding defense design against potential countersample attacks.
The countermeasure sample detection technology is a countermeasure sample defense method which does not affect the structure and the parameters of the recognition model. In the face of the confrontation sample image which has no visual difference with the normal image, the confrontation sample detection technology can find the difference between the normal sample and the confrontation sample and send out early warning in advance. Because the method has the characteristic of not influencing the model precision, the countercheck sample detection technology is very suitable for being applied to an SAR recognition system requiring the stable model effect. However, due to the characteristics of the SAR image, the implementation of the SAR challenge sample detection task becomes abnormally difficult, for example, speckle noises which are difficult to eliminate are distributed in the SAR image, and the noises of the challenge sample are mixed in the image, so that the model is misled to make a wrong classification result. The existing traditional countermeasure sample detection method based on kernel density estimation, bayesian estimation and inherent dimension is difficult to obtain the characteristics for effectively distinguishing the countermeasure samples in the model interior. In addition, the SAR image has higher resolution than the traditional image, which causes the internal intermediate features of the model to become extremely large, and the existing analysis method based on the model intermediate features applied to the field of the traditional image is difficult to process the large number of features. In addition, targets in different application scenes have different potential characteristics, and the existing analysis method is difficult to automatically adjust according to task targets. Therefore, research and development of a detection technology capable of performing automatic adjustment according to a scene, processing a high-resolution SAR image, and effectively distinguishing the characteristics of a challenge sample is urgently needed.
Disclosure of Invention
The invention aims to provide a system and a method for detecting SAR image confrontation samples based on multi-target optimal characteristic cause selection aiming at the defects of the prior art.
The purpose of the invention is realized by the following technical scheme: an SAR image confrontation sample detection system based on multi-target optimal feature attribution selection comprises an SAR image confrontation sample detection data preprocessing module, an SAR image confrontation sample detection off-line optimal feature attribution selection training module and an SAR image confrontation sample detection on-line detection module;
the SAR image confrontation sample detection data preprocessing module acquires real-time image data in a monitoring process from a real-time database of a target SAR system, obtains an online detection data set after image data normalization and image data normalization processing, and transmits the online detection data set to the SAR image confrontation sample detection online detection module; the SAR image confrontation sample detection data preprocessing module acquires historical image data from a historical database of a target SAR system, obtains standard historical image data after image data normalization and image data normalization, generates confrontation sample image data corresponding to the standard historical image data, combines the standard historical image data and the confrontation sample image data to obtain an offline training data set, and transmits the offline training data set to an SAR image confrontation sample detection offline optimal characteristic cause selection training module;
the SAR image countermeasure sample detection offline optimal feature attribution selection training module is characterized by firstly carrying out integer coding on the Size (Size) in a feature attribution scanning block, the scanning interval (Stride), a mark (tagging) for whether expansion operation is carried out or not when the edge of an image is scanned, the number (Layers) of model hidden Layers for feature analysis, randomly generating an initialization population, then carrying out sliding scanning on SAR image data in an offline training data set based on the feature attribution scanning blocks corresponding to different individual codes through a sliding scanning submodule, carrying out feature extraction through a hidden feature extraction submodule, calculating a feature expression set corresponding to the offline training data set through a feature expression calculation submodule, carrying out logistic regression training by taking the obtained feature expression set of the offline training data set as training data through a regression model training submodule, and obtaining a logistic regression model for identifying the countermeasure sample; counting the number of sub-samples generated in the characteristic analysis process and the Area (AUC) Under the working characteristic Curve of a subject of the regression model as the fitness index of the individual fitness function evaluation submodule, sequencing the individuals in the initial population by adopting a quick non-dominated sequencing and fitness comprehensive evaluation submodule to obtain pareto frontier individuals, and selecting the optimal individual from the pareto frontier individuals. Generating offspring populations through selection, crossing and mutation, merging the parents and the offspring to generate a new population, performing rapid non-dominated sorting and crowding distance calculation on the new population, and generating a next generation population according to the population scale; repeating the above evolution process until a maximum number of evolutionary rounds is reached, thereby obtaining individuals with pareto optima; transmitting the optimal characteristic attribution scanning block and the optimal regression model corresponding to the optimal individual to an SAR image confrontation sample detection online detection module;
the SAR image countermeasure sample detection online detection module performs sliding scanning on SAR image data in an online detection data set by using an optimal characteristic cause scanning block through a sliding scanning submodule, performs characteristic extraction through an implicit characteristic extraction submodule, and calculates a characteristic expression set of the online detection data set through a characteristic expression calculation submodule; and then using the obtained optimal regression model as a challenge sample detection model to judge whether the feature expression of the online detection data set belongs to the feature expression of the challenge sample. If detecting that a certain characteristic expression belongs to the characteristic expression of the confrontation sample, sending out early warning information; otherwise, the sample is a normal sample;
the SAR image confrontation sample detection method based on multi-target optimal feature attribution selection comprises the following steps:
(1) The SAR image confrontation sample detection data preprocessing module acquires historical SAR image data stored in a monitoring process from an SAR system as an original data set, and marks the historical SAR image data as an O; carrying out image data normalization operation on the O to obtain a normalized data set X, and then carrying out image data normalization operation through a formula (1) to obtain an offline standard data set X o ；
The SAR image data normalization module consists of segmentation operation, scaling operation and dimension reduction operation. The segmentation operation means that the original image is segmented into regions, and when the input SAR historical monitoring image is a monitoring region overall image containing a plurality of identification targets, the image needs to be segmented according to the identification targets, so that the segmented SAR image only contains a single identification target; when the input SAR historical image data is segmented and each SAR historical image only contains a single identification target, the segmentation operation is not needed; scaling operation means that the size of each SAR image in the data set is adjusted, and the size parameter of each SAR image is ensured to be consistent with the parameter of the target model; the dimensionality reduction operation represents that single-channel processing is carried out on the SAR image in the data set; if the input SAR historical image data is a single-channel image, the dimension reduction operation is not needed;
will take off-line standard data set X o Transmitting to a challenge sample generation platform for challenge sample generation, i.e. using 5 challenge sample generation methods for generating X o Confrontational sample data set X under target model adv (ii) a The 5 challenge sample generation methods include: FGSM (Fast Gradient Signal Method) attack based on Fast Gradient Sign, PGD (Project Gradient Description) attack based on Gradient iteration, CW (Carlini) based on optimization&Wagner) attack, boundary decision-based Deepfool attack, and random Noise-based Noise attack; mixing X o And X adv And (3) segmenting according to the proportion of 4 o_train Training set X containing challenge samples adv_train Verification set X of clean data o_val Validation set X containing challenge samples adv_val ；
(2) Setting relevant parameter values in an offline optimal feature attribution selection training module, wherein the parameter values comprise a population Size N and a maximum Size of a feature attribution scanning block max Minimum Size min Maximum step size Stride max Minimum step size Stride min Maximum number of layers L selected for a feature layer, training round EP train Cross rate beta, variance rate sigma, maximum evolution round E max ；
(3) Performing integer coding on a set of parameters of Size, stride, padding and Layers in a feature attributed scanning block, as an individual in a multi-objective evolutionary strategy, initializing N individuals as an initial population Q, wherein the coding form of each individual is Indi = [ Size, stride, padding, layers ], indi represents any one individual in the population, wherein Size represents the Size of the feature attributed scanning block, stride represents the scanning interval of the feature attributed scanning block, padding represents a mark of whether the feature attributed scanning block performs an expansion operation when scanning an image edge, layers represent the number of model hidden Layers for feature analysis, and the implementation processes of Size, stride, padding and layer initialization are respectively shown in formulas (2) to (5):
Size＝Randint(Size min ,Size max )
(18)
Stride＝Randint(Stride min ,Stride max ) (19)
Padding＝Randint(0,1) (20)
Layers＝Randint(1,L) (21)
the random (a, b) represents that an integer with the size between a and b is randomly generated, and the value range comprises two end values of a and b; when Padding =0, the feature scanning process does not perform the edge extension operation, and when Padding =1, the feature scanning process performs the edge extension operation; mixing X o The length and the width of the medium SAR image are respectively marked as h and w, h = w under the general condition, the line scanning times of the characteristic cause scanning block is marked as m, and the calculation process of m is shown as a formula (6); the edge pixel width not covered by the line scanning process is marked as k, and the calculation process of k is shown in formula (7):
k＝(h-Size)-m×Stride (23)
wherein
Represents a round-down operation; when Padding =0, k is not equal to 0, it represents that the current Size and Stride parameter values cannot scan the function edge, and part of edge pixels of the SAR image do not participate in the subsequent feature extraction process; when Padding =1 and k is not equal to 0, performing edge expansion on the SAR image before feature attribution analysis, wherein the value of a filled pixel value is fixed to 0, so that all pixels in the SAR image participate in the subsequent feature attribution analysis process;
the hidden feature extraction operation can be used for sequentially extracting the output features of the hidden Layers from the last hidden layer of the target model according to the number of Layers until the number of the extracted hidden Layers is equal to the number of Layers;
(4) Marking the current evolution turn of the multi-objective optimization technology as E, and enabling E =0;
(5) Evaluating a fitness function of the initial population Q, namely, respectively calculating a data set X by an SAR image confrontation sample detection offline optimal characteristic attribution selection training module according to characteristic attribution scanning blocks corresponding to individuals in Q o_train ，X adv_train ，X o_val ，X adv_val The characteristic expression of (a); then, using training set X o_train ，X adv_train Performing logistic regression training on the characteristic expression set with the training turn being EP train Using verification set X in the training process o_val ，X adv_val The obtained regression model is verified by the characteristic expression set, the AUC value of the regression model is calculated, and the number of the sub-samples generated in the characteristic analysis process is evaluated;
the specific calculation process of the single data set characteristic attribution expression is as follows:
(5.1) let i =0, label the maximum number of images of the dataset as i max ；
(5.2) setting a set F of subsample images sub Selecting the ith image x of the data set as a current feature analysis image for an empty set, and performing sliding scanning on the image by using a feature attribution scanning block; starting from the (0, 0) position in the x image pixel matrix, selecting a square area with the Size of Size in the x image pixel matrix according to the Size parameter to perform sub-sample generation operation; the sub-sample generation operation clones a new image with the same value as the original x pixel value, sets the pixel value of the scanning area of the new image to 0, and then adds the image into the sub-sample image set F sub Performing the following steps; after one-time sub-sample generation operation is completed, the characteristic attribution scanning block slides in rows according to the Stride parameters, and one-time sub-sample generation operation is performed in each sliding scanning; repeating the process until the next scanning of the scanning block exceeds the width of the pixel matrix, moving the scanning block to the next row scanning position in the x image pixel matrix according to the Stride parameters, and starting to slide by rows againAnd (4) moving scanning. Repeating the above process until the feature cause scanning block completes all the line-by-line scanning operations in the x matrix pixels; finally, the original image x is also added into the sub-sample image set F sub Performing the following steps; the starting point position of the p-th scanning of the l-th line in the sliding scanning operation can be recorded as (l × Stride, p × Stride), and the coordinates of the four points of the corresponding square scanning area are respectively: (l × Stride, p × Stride), (l × Stride, p × Stride + Size), (l × Stride + Size, p × Stride + Size);
(5.3) set F of subsamples of the image sub Inputting a target model, selecting a corresponding model hidden layer according to a given Layers parameter, and obtaining the output of each image in the subsample set corresponding to the hidden layer in the model; carrying out average pooling operation on the output of the hidden layer to obtain a subsample set F sub A corresponding set of implicit features;
(5.4) subtracting the implicit characteristic corresponding to the original image x from the implicit characteristic corresponding to the sub-sample image except the original image x in the implicit characteristic set to obtain an image characteristic change matrix capable of representing the influence of image pixel change;
(5.5) carrying out quartile distance (IQR) statistics on the image characteristic change matrix to obtain a characteristic expression vector capable of representing the image;
(5.6) let i = i +1;
(5.7) repeating (5.2) to (5.6) steps until i = i max ；
And (5.8) merging all the feature expression vectors obtained in the step (5.7) to obtain a feature expression set corresponding to the data set.
The specific implementation process of the AUC value evaluation is as follows:
verification set X o_val The number of samples in (1) is R, and the samples are marked as positive samples; verification set X adv_val The number of samples in (1) is T, and the label is negative. The prediction score of the verification set sample in the regression model is marked as P, P Positive sample Representing the prediction score, P, of the regression model on a single positive sample Negative sample Representing the predicted score of the regression model for a single negative sample; note I (P) Positive sample ,P Negative sample ) Predicting an evaluation value for a sample of a positive and negative sample pair; when P is present Positive sample >P Negative sample When is, I (P) Positive sample ,P Negative sample ) =1; when P is Positive sample ＝P Negative sample When, I (P) Positive sample ,P Negative sample ) =0.5; when P is present Positive sample <P Negative sample When is, I (P) Positive sample ,P Negative sample ) =0; calculating the sample prediction evaluation value of all the positive and negative sample pairs, and calculating the AUC value of the regression model to the verification set sample according to the formula (8):
the specific implementation process of calculating the number of the sub-samples generated in the feature analysis process is as follows: using the feature attribution scan block parameters Size, stride, padding corresponding to the individuals, calculating k and m corresponding to the current individuals through formula (6) and formula (7); when Padding =0, the number of individually corresponding subsamples is m × m +1; when Padding =1, if k =0, the individually corresponding number of subsamples is m × m +1; when Padding =1, if k ≠ 0, the number of individually corresponding subsamples is (m + 1) × (m + 1) +1;
(6) And (4) carrying out rapid non-dominated sorting and comprehensive fitness evaluation on individuals in the population Q. The fast non-dominant ranking searches the pareto frontier of the population Q by measuring the advantages and disadvantages of the individuals in two evaluation indexes of an AUC value and the number of sub samples, and the fitness comprehensively evaluates the two evaluation indexes in the pareto frontier individual set to select the optimal individual of the population Q. In specific operation, the negative number and the sub-sample number of the AUC value of an individual are taken as optimization objective functions and respectively marked as f 1 And f 2 (ii) a Note n i To determine the number of individuals in the population that dominate the ith individual, S i A set of individuals dominated by the ith individual; what an individual dominates over another is: for the AUC value corresponding to the individual and the number of subsamples, the dominant individual performs better than the dominated individual, i.e. the model AUC value is higher than the dominated individual, and the number of subsamples is smaller than the dominated individual. Using fast non-dominant rowsObtaining a dominant individual set, and selecting an optimal individual by combining fitness comprehensive evaluation, wherein the specific implementation process comprises the following steps:
(6.1) traversing all individuals of the current population Q, and calculating n corresponding to each individual i Value and the number governed by the individual
Body set S i ；
(6.2) mixing all n i Individuals of =0 are stored in the set F 1 Performing the following steps;
(6.3) let j =1;
(6.4) letting H be an empty set;
(6.5) traverse F j All individuals in (1), note S u Is as a quilt F j The set of individuals governed by the u-th individual, n q To dominate S u Number of subjects in qth; to obtain F j S corresponding to each individual in u ；
(6.6) for all S u Go through each S u Calculating n corresponding to each individual q A value of and let n q ＝n q -1；
(6.7) if n q If not less than 0, then S u The q-th individual in (a) is placed in set H;
(6.8) let j = j +1;
(6.9) order F j ＝H；
(6.10) repeating (6.4) to (6.9) steps until F is obtained j Is an empty set;
(6.11)F 1 all of them are dominant ones, and F is 1 The individual in (1) is used as a pareto frontier individual of a population Q, the optimal individual is selected by integrating two evaluation indexes, and the specific selection process is as follows: when F is 1 Only contains 1 individual, the individual is selected as the optimal individual Indi in the evaluation best (ii) a When F is present 1 2 individuals, the individual with the higher AUC value is selected as the optimal individual Indi in the evaluation best (ii) a When F is present 1 When 3 or more than 3 individuals are included in the evaluation data, the fitness comprehensive evaluation value f of the ith individual is calculated according to the formula (9) i all Wherein f is 1 i F representing the ith individual 1 Optimization of the target value, f 2 i F representing the ith individual 2 The target value is optimized. Selection f i all The smallest individual is regarded as the optimal individual, indi best When there are multiple individuals with the smallest f at the same time i all Then, the individuals with higher AUC values are selected as the optimal individual Indi of the model best ；
(7) For each individual in the population Q, a binary competitive competition selection method is adopted, namely pairwise comparison is carried out, and each comparison is carried out according to the optimal individual evaluation method in the step (6) to select the individual with the smaller f i all The individuals with the value enter the parent population Q p ；
(8) For population Q p Each individual in the group is subjected to a crossover operation by using an analog binary crossover method. Marking the crossed population as Q x The concrete implementation process is as follows:
(8.1) generating a uniformly distributed random number β in the range of 0 to 1 for the current individual 1 When is beta 1 When the beta is less than or equal to beta, the current individual and another random individual are taken as parents to carry out subsequent operation, and when the beta is less than or equal to beta, the current individual and another random individual are taken as parents 1 >Beta, no operation is carried out;
(8.2) generating a uniformly distributed random number g in the range of 0 to 1 1 Remember k 1 The same coding position for two individuals in the cross operation;
(8.3) calculating the crossover intermediate variable β according to equation (10) g1 ；
(8.4) calculating the parent influence factor beta according to the formula (11) k1 ；
(8.5) Indi for two individuals 1 And Indi 2 The two filial generation individuals generated by the crossover operation are Indi c1 And Indi c2 。Indi c1 K th of (1) 1 The value of each position is calculated by the formula (12), indi c2 K of (2) 1 The value of each position is calculated according to equation (13). Wherein Indi c1 (k 1 ) Is Indi c1 K of (2) 1 Value of individual position, indi 1 (k 1 ) Is Indi 1 K of (2) 1 The value of each position, the rest is the same;
(8.6) adding Indi c1 And Indi c2 The non-integer values are rounded to integers;
(8.7) repeating (8.1) to (8.6) steps until Q p Each individual in (a) performs a crossover operation. Putting all the obtained cross-offspring individuals into a population Q x Performing the following steps;
(9) For population Q x Performing mutation operation by a polynomial mutation method, and marking the mutated population as Q m . The specific implementation process is as follows:
(9.1) generating a uniformly distributed random number σ in the range of 0 to 1 for the current individual 1 When σ is 1 When the sigma is less than or equal to the sigma, the subsequent operation is carried out on the individual, and when the sigma is less than or equal to the sigma 1 >When sigma is, no operation is carried out;
(9.2) generating a uniformly distributed random number g in the range of 0 to 1 2 Remember k 2 The coding position of an individual in the mutation operation;
(9.3) calculating the intermediate variable of variation beta according to equation (14) g2 ；
(9.4) recording k of the individual Indi 2 The upper bound of the value of the position is Indi (k) 2 ) up K of the individual Indi 2 The lower bound of the value of position is Indi (k) 2 ) low . The offspring generated by the mutation recording operation is Indi mut ，Indi mut K of (2) 2 The value of each position is calculated by formula (15);
Indi mut (k 2 )＝Indi(k 2 )+(Indi(k 2 ) up -Indi(k 2 ) low )×β g2 (31)
(9.5) adding Indi mut The non-integer values are rounded to integers;
(9.6) repeating (9.1) to (9.5) steps until Q x Was subjected to mutation operation. Putting all obtained variant progeny individuals into a population Q m Performing the following steps;
(10) The population Q p And child population Q x , child population Q m Are combined into a new population Q n ；
(11) For new population Q n And (4) implementing an elite strategy to select and generate a next generation population. Q n Comprises an individual set Q of a previous generation population after being selected by a binary competitive bidding competition p I.e. including the best individual Indi in the previous generation population best . Elite strategy will be the best individual Indi of the previous generation best And carrying out fitness evaluation and other operations with the new filial generation individuals to generate a next generation population. Recording the next generation population as Q p+1 The detailed process of the elite selection strategy is as follows:
(11.1) evaluating the new population Q according to the steps (5) and (6) n All individuals in (1) perform a fast non-dominated sorting operation to obtain Q n All non-dominant sets F of j 。
(11.2) let Q p+1 As empty set, j =1;
(11.3) calculation of F j The crowdedness distance of all individuals in the group. Note d i Is F j The crowdedness distance of the ith individual is f 1 And f 2 To F j Will have the largest f 1 Value and maximum f 2 Two individuals of the value are used as boundary individuals, and the degree of crowding d corresponding to the boundary individuals i Considered infinite. The crowdedness distances of the other individuals except the boundary individual are calculated according to the formula (16). Wherein f is z i+1 Is represented by F j Of (i + 1) th individual z Target fitness value, f z i-1 Is represented by F j Of (i-1) th individual z A target fitness value;
(11.4) recording of | Q p+1 Is Q | p+1 Number of individuals, | F j L is F j Number of individuals in the population. When | Q p+1 |+|F j When | < N, adding F j All individuals in (1) put in Q p+1 In (1). When | Q p+1 |+|F j |>When N is present, F is j The middle individuals are sorted from large to small according to the crowdedness distance, and F is arranged after sorting j In the sequence, N- | Q is selected p+1 | Individual put in Q p+1 Performing the following steps;
(11.5) let j = j +1;
(11.6) repeating (11.1) to (11.5) steps until | Q p+1 |＝N；
(12) Let Q = Q p+1 ，E＝E+1；
(13) Repeating the steps (5) to (12) until E = E max ；
(14) For the passage E max And (5) performing rapid non-dominated sorting and comprehensive fitness evaluation on the population Q subjected to the round evolution according to the modes of the step (5) and the step (6) to obtain the optimal individual Indi best . The optimal individual index best The corresponding characteristic cause scanning block and the logistic regression model are transmitted to an SAR image confrontation sample detection online detection module;
(15) SAR image countermeasure sample detectionThe measured data preprocessing module acquires real-time image data in the monitoring process from the target SAR system, and an online detection data set X is obtained after data normalization and data normalization processing t The online detection data set X is t Transmitting the data to an online detection module;
(16) The SAR image confrontation sample detection online detection module uses the obtained optimal characteristic attribution scanning block to calculate an online detection data set X according to the characteristic attribution expression calculation mode in the step (5) t The characteristic expression of (a); then judging X by using the obtained optimal regression model t Whether the characteristic expression of the countermeasure sample belongs to the characteristic expression of the countermeasure sample or not and sending out early warning on the found characteristic expression of the countermeasure sample; otherwise, the sample is a normal sample.
The invention has the beneficial effects that: the technology of the invention can automatically obtain the optimal feature analysis granularity according to different scenes, and efficiently realize the detection of various confrontation samples in the SAR image recognition field; compared with the prior art, the SAR image confrontation detection method and device further improve the calculation efficiency and AUC performance indexes of SAR image confrontation detection.
Drawings
FIG. 1 is a schematic diagram of a SAR image countermeasure sample detection system and method based on multi-objective optimal feature cause selection;
fig. 2 is an exemplary diagram of a randomly generated single individual code. The coded information of the individual is [34,15,1,2], that is, the Size of the feature attribution scanning block corresponding to the individual is 34 pixels (Size = 34), the scanning interval is 15 pixels (Stride = 15), the Padding operation is performed when the image edge is scanned (Padding = 1) and the number of model hidden Layers adopted when the feature attribution scanning block is used for feature analysis is 2 (Layers = 2);
FIG. 3 is an exemplary diagram of interleaving operations in a feature scan block parameter optimization process;
FIG. 4 is an exemplary diagram of mutation operations in a feature scan block parameter optimization process;
FIG. 5 is a final pareto front obtained through a multi-objective optimization technique;
FIG. 6 is a comparison of AUC performance of the present invention and existing challenge sample detection techniques against the FUSAR-Ship dataset over 5 challenge sample types. KDBU represents a confrontation sample detection technology based on nuclear density and Bayesian estimation, LID represents a confrontation sample detection technology based on intrinsic dimension, and MAHA represents a confrontation sample detection technology based on a Mahalanobis distance score;
FIG. 7 is a comparison of the present invention and the pixel Leave-One-Out (LOO) based countermeasure sample detection technique on the number of subsamples. Compared with the three existing antagonistic sample detection technologies, the antagonistic sample detection technology based on the LOO sacrifices the operation efficiency to obtain a better detection effect. In a specific implementation process, the technology gradually modifies a single pixel in an input sample to generate a sub-sample, analyzes an obtained sub-sample set, and calculates a feature expression corresponding to the sample. When the size of the input sample is large, the implementation time of the technology can be obviously increased;
Detailed Description
The purpose and effect of the present invention will be more apparent from the following further description of the present invention with reference to the accompanying drawings.
By taking the FUSAR-Ship image dataset as an example of detecting the SAR image confrontation samples, a schematic diagram of the SAR image confrontation sample detection system and method based on multi-target optimal feature cause selection is given in FIG. 1. The SAR image confrontation sample detection system based on multi-target optimal characteristic cause selection comprises an SAR image confrontation sample detection data preprocessing module, an SAR image confrontation sample detection off-line optimal characteristic cause selection training module and an SAR image confrontation sample detection on-line detection module;
the SAR image confrontation sample detection data preprocessing module acquires real-time image data in a monitoring process from a real-time database of a target SAR system, obtains an online detection data set after image data normalization and image data normalization processing, and transmits the online detection data set to the SAR image confrontation sample detection online detection module; the SAR image confrontation sample detection data preprocessing module acquires historical image data from a historical database of a target SAR system, obtains standard historical image data after image data normalization and image data normalization, generates confrontation sample image data corresponding to the standard historical image data, combines the standard historical image data and the confrontation sample image data to obtain an offline training data set, and transmits the offline training data set to an SAR image confrontation sample detection offline optimal characteristic cause selection training module;
the SAR image countermeasure sample detection offline optimal feature attribution selection training module is characterized by firstly carrying out integer coding on the Size (Size) in a feature attribution scanning block, the scanning interval (Stride), a mark (tagging) for whether expansion operation is carried out or not when the edge of an image is scanned, the number (Layers) of model hidden Layers for feature analysis, randomly generating an initialization population, then carrying out sliding scanning on SAR image data in an offline training data set based on the feature attribution scanning blocks corresponding to different individual codes through a sliding scanning submodule, carrying out feature extraction through a hidden feature extraction submodule, calculating a feature expression set corresponding to the offline training data set through a feature expression calculation submodule, carrying out logistic regression training by taking the obtained feature expression set of the offline training data set as training data through a regression model training submodule, and obtaining a logistic regression model for identifying the countermeasure sample; and (3) counting the number of sub samples generated in the characteristic analysis process and the Area Under the working characteristic Curve (AUC) of a subject of the regression model as the fitness index of the individual fitness function evaluation submodule, sequencing the individuals in the initial population by adopting a quick non-dominated sequencing and fitness comprehensive evaluation submodule to obtain pareto front individuals, and selecting the optimal individual from the pareto front individuals. Generating offspring populations through selection, crossing and mutation, merging the parents and the offspring to generate a new population, performing rapid non-dominated sorting and crowding distance calculation on the new population, and generating a next generation population according to the population scale; repeating the above evolution process until a maximum number of evolutionary rounds is reached, thereby obtaining individuals with pareto optima; transmitting the optimal characteristic cause scanning block and the optimal regression model corresponding to the optimal individual to an SAR image confrontation sample detection online detection module;
the SAR image confrontation sample detection online detection module performs sliding scanning on SAR image data in an online detection data set by using an optimal characteristic attribution scanning block through a sliding scanning submodule, performs characteristic extraction through an implicit characteristic extraction submodule and calculates a characteristic expression set of the online detection data set through a characteristic expression calculation submodule; and then using the obtained optimal regression model as a challenge sample detection model to judge whether the feature expression of the online detection data set belongs to the feature expression of the challenge sample. If detecting that a certain characteristic expression belongs to the characteristic expression of the countermeasure sample, sending out early warning information; otherwise, the sample is a normal sample;
the SAR image confrontation sample detection method based on multi-target optimal characteristic cause selection comprises the following steps:
(1) The SAR image confrontation sample detection data preprocessing module acquires historical SAR image data stored in a monitoring process from an SAR system as an original data set, and marks the historical SAR image data as an O; carrying out image data normalization operation on the O to obtain a normalized data set X, and then carrying out image data normalization operation through a formula (1) to obtain an offline standard data set X o ；
The SAR image data normalization module consists of segmentation operation, scaling operation and dimension reduction operation. The segmentation operation means that the original image is segmented into regions, when the input SAR historical monitoring image is a monitoring region overall image containing a plurality of identification targets, the image needs to be segmented according to the identification targets, and the segmented SAR image only contains a single identification target; when the input SAR historical image data is segmented and each SAR historical image only contains a single identification target, the segmentation operation is not needed; scaling operation means that the size of each SAR image in the data set is adjusted, and the size parameter of each SAR image is ensured to be consistent with the parameter of the target model; the dimensionality reduction operation represents that single-channel processing is carried out on the SAR image in the data set; if the input SAR historical image data is a single-channel image, the dimension reduction operation is not needed;
off-line standard data set X o Transmitting to a challenge sample generation platform for challenge sample generation, namely using 5 challenge sample generation methods for generating X o Countermeasure sample data set X under target model adv (ii) a The 5 challenge sample generation methods include: FGSM (Fast Gradient Signal Method) attack based on Fast Gradient Sign, PGD (Project Gradient Description) attack based on Gradient iteration, CW (Carlini) based on optimization&Wagner) attack, boundary decision-based Deepfool attack, and random Noise-based Noise attack; mixing X o And X adv And (3) segmenting according to the proportion of 4 o_train Training set X containing challenge samples adv_train Verification set X of clean data o_val Validation set X comprising challenge samples adv_val ；
(2) Setting relevant parameter values in an offline optimal feature attribution selection training module, wherein the parameter values comprise a population Size N =20 and a maximum Size of a feature attribution scanning block max =64, minimum Size min =8, maximum step size Stride max =64, minimum step size Stride min =8, maximum number of layers selected for the eigen layer L =3, training round EP train =100, crossover ratio β =1.0, variability ratio σ =0.5, maximum round of evolution E max ＝20；
(3) Performing integer coding on a set of parameters of Size, stride, padding and Layers in a feature attributed scanning block, as an individual in a multi-objective evolutionary strategy, initializing N individuals as an initial population Q, wherein the coding form of each individual is Indi = [ Size, stride, padding, layers ], indi represents any one individual in the population, wherein Size represents the Size of the feature attributed scanning block, stride represents the scanning interval of the feature attributed scanning block, padding represents a mark of whether the feature attributed scanning block performs an expansion operation when scanning an image edge, layers represent the number of model hidden Layers for feature analysis, and the implementation processes of Size, stride, padding and layer initialization are respectively shown in formulas (2) to (5):
Size＝Randint(Size min ,Size max )
(34)
Stride＝Randint(Stride min ,Stride max ) (35)
Padding＝Randint(0,1) (36)
Layers＝Randint(1,L) (37)
the random (a, b) represents that an integer with the size between a and b is randomly generated, and the value range comprises two end values of a and b; when Padding =0, the feature scanning process does not perform the edge extension operation, and when Padding =1, the feature scanning process performs the edge extension operation; mixing X o The length and the width of the medium SAR image are respectively marked as h and w, h = w under the general condition, the line scanning times of the characteristic cause scanning block is marked as m, and the calculation process of m is shown as a formula (6); the edge pixel width not covered by the line scanning process is marked as k, and the calculation process of k is shown in formula (7):
k＝(h-Size)-m×Stride (39)
wherein
Represents a round-down operation; when Padding =0, k is not equal to 0, it represents that the current Size and Stride parameter values cannot scan the function edge, and part of edge pixels of the SAR image do not participate in the subsequent feature extraction process; when Padding =1, k is not equal to 0, performing edge expansion on the SAR image before feature attribution analysis, and fixing the value of the filled pixel value to be 0, so that all pixels in the SAR image participate in the subsequent feature attribution analysis process;
and the implicit feature extraction operation can be used for sequentially extracting the output features of the implicit Layers from the last implicit layer of the target model according to the number of Layers until the number of the extracted implicit Layers is equal to the value of Layers. FIG. 2 shows an example of the encoding of one of the initial individuals whose encoded information is [34,15,1,2];
(4) Marking the current evolution turn of the multi-objective optimization technology as E, and enabling E =0;
(5) Evaluating a fitness function of the initial population Q, namely, respectively calculating a data set X by an SAR image confrontation sample detection offline optimal characteristic attribution selection training module according to characteristic attribution scanning blocks corresponding to individuals in Q o_train ，X adv_train ，X o_val ，X adv_val The characteristic expression of (a); then, using training set X o_train ，X adv_train Performing logistic regression training on the feature expression set, wherein the training turn is EP train Using verification set X in the training process o_val ，X adv_val The obtained regression model is verified by the characteristic expression set, the AUC value of the regression model is calculated, and the number of the sub-samples generated in the characteristic analysis process is evaluated;
the specific calculation process of the single data set characteristic attribution expression is as follows:
(5.1) let i =0, label the maximum number of images of the dataset as i max ；
(5.2) set the subsample image set F sub Selecting the ith image x of the data set as a current feature analysis image for an empty set, and performing sliding scanning on the image by using a feature attribution scanning block; starting from the (0, 0) position in the x image pixel matrix, the scanning operation selects a square area with the Size of Size in the x image pixel matrix according to the Size parameter to perform sub-sample generation operation; the sub-sample generation operation clones a new image with the same value as the original x pixel value, sets the pixel value of the scanning area of the new image to 0, and then adds the image into the sub-sample image set F sub Performing the following steps; after one-time sub-sample generation operation is completed, the characteristic attribution scanning block slides in rows according to the Stride parameters, and one-time sub-sample generation operation is performed in each sliding scanning; repeating the process until the next scan of the scan block exceeds the width of the pixel matrix, moving the scan block to the next row scan position in the x image pixel matrix according to the Stride parameters, and starting the line-by-line sliding scan again. Repeating the above process until the feature cause scanning block completes all the line-by-line scanning operations in the x matrix pixels; finally, the original image x is also added into the sub-sample image set F sub Performing the following steps; the starting point position of the p-th scanning of the l-th line in the sliding scanning operation can be recorded as (l × Stride, p × Stride), and the coordinates of the four points of the corresponding square scanning area are respectively: (l × Stride, p × Stride), (l × Stride, p × Stride + Size), (l × Stride + Size, p × Stride + Size, and Size). In units of [34,15,1,2]For example, when the image is slide-scanned by using the feature-attributed scanning block corresponding to the individual, the starting point position of the p-th scanning of the ith row can be written as (15 × l,15 × p), and the four point coordinates of the corresponding square scanning area are: (15. Times.l, 15. Times.p), (15. Times.l, 15. Times.p + 34), (15. Times.l +34, 15. Times.p + 34);
(5.3) set F of subsamples of the image sub Inputting a target model, selecting a corresponding model hidden layer according to a given Layers parameter, and obtaining the output of each image in the subsample set corresponding to the hidden layer in the model; carrying out average pooling operation on the output of the hidden layer to obtain a subsample set F sub A corresponding set of implicit features;
(5.4) subtracting the implicit features corresponding to the original image x from the implicit features corresponding to the sub-sample images except the original image x in the implicit feature set to obtain an image feature change matrix capable of representing the influence of image pixel change;
(5.5) carrying out quartile distance (IQR) statistics on the image characteristic change matrix to obtain a characteristic expression vector capable of representing the image;
(5.6) let i = i +1;
(5.7) repeating (5.2) to (5.6) steps until i = i max ；
And (5.8) merging all the characteristic expression vectors obtained in the step (5.7) to obtain a characteristic expression set corresponding to the data set.
The specific implementation process of the AUC value evaluation is as follows:
verification set X o_val The number of samples in (1) is R, and the samples are marked as positive samples; verification set X adv_val The number of samples in (1) is T, and the label is negative. The prediction score of the verification set sample in the regression model is marked as P, P Positive sample Representing the prediction score, P, of the regression model on a single positive sample Negative sample Representing the predicted score of the regression model for a single negative sample; note I (P) Positive sample ,P Negative sample ) Predicting an evaluation value for a sample of a positive and negative sample pair; when P is present Positive sample >P Negative sample When is, I (P) Positive sample ,P Negative sample ) =1; when P is present Positive sample ＝P Negative sample When is, I (P) Positive sample ,P Negative sample ) =0.5; when P is present Positive sample <P Negative sample When is, I (P) Positive sample ,P Negative sample ) =0; calculating the sample prediction evaluation value of all the positive and negative sample pairs, and calculating the AUC value of the regression model to the verification set sample according to the formula (8):
the specific implementation process of calculating the number of the sub-samples generated in the feature analysis process is as follows: attributing scan block parameters Size, stride, padding using features corresponding to individuals, and calculating k and m corresponding to the current individual through formula (6) and formula (7); when Padding =0, the number of individually corresponding subsamples is m × m +1; when Padding =1, if k =0, the individually corresponding number of subsamples is m × m +1; when Padding =1, if k ≠ 0, the number of subsamples corresponding to the individual is (m + 1) × (m + 1) +1. Taking an individual [34,15,1,2] as an example, padding =1, and m =31,k =13 is calculated, the number of corresponding subsamples of the individual is (31 + 1) × (31 + 1) +1=1025;
(6) And (4) carrying out rapid non-dominated sorting and comprehensive fitness evaluation on the individuals in the population Q. And the fast non-dominated sorting is used for searching the pareto frontier of the population Q by measuring the advantages and disadvantages of the individuals in AUC values and two evaluation indexes of the number of sub-samples, and the fitness comprehensive evaluation is used for comprehensively evaluating the two evaluation indexes in the pareto frontier individual set to select the optimal individual of the population Q. In specific operation, the negative number and the sub-sample number of the AUC value of an individual are used as an optimization objective functionAre respectively marked as f 1 And f 2 (ii) a Note n i Number of individuals dominating the ith individual in the population, S i A set of individuals dominated by the ith individual; the meaning that one individual dominates another is: for the AUC value corresponding to the individual and the number of subsamples, the dominant individual performs better than the dominated individual, i.e. the model AUC value is higher than the dominated individual, and the number of subsamples is smaller than the dominated individual. The rapid non-dominated sorting is adopted to obtain an individual set in a dominated position, and an optimal individual is selected by combining fitness comprehensive evaluation, and the specific implementation process is as follows:
(6.1) traversing all individuals of the current population Q, and calculating n corresponding to each individual i Value and the number governed by the individual
Body set S i ；
(6.2) mixing all n i The individuals of =0 are stored in the set F 1 The preparation method comprises the following steps of (1) performing;
(6.3) let j =1;
(6.4) making H as an empty set;
(6.5) traverse F j All individuals in (1), note S u Is as a quilt F j The set of individuals governed by the u-th individual, n q To dominate S u Number of subjects in qth; to obtain F j S corresponding to each individual in u ；
(6.6) for all S u Go through each S u Calculating n corresponding to each individual q A value of and let n q ＝n q -1；
(6.7) if n q If not less than 0, then S u The q-th individual in (a) is placed in set H;
(6.8) let j = j +1;
(6.9) order F j ＝H；
(6.10) repeating (6.4) to (6.9) steps until F is obtained j Is an empty set;
(6.11)F 1 all of them are dominant ones, and F is 1 The individuals in the population Q are used as pareto frontier individuals of the population Q, the optimal individuals are selected by integrating two evaluation indexes,the specific selection process comprises the following steps: when F is present 1 Only contains 1 individual, the individual is selected as the optimal individual Indi in the evaluation best (ii) a When F is present 1 2 individuals, the individual with the higher AUC value is selected as the optimal individual Indi in the evaluation best (ii) a When F is present 1 When 3 or more than 3 individuals are included in the evaluation data, the fitness comprehensive evaluation value f of the ith individual is calculated according to the formula (9) i all Wherein f is 1 i F representing the ith individual 1 Optimization of the target value, f 2 i F representing the ith individual 2 The target value is optimized. Selection f i all The smallest individual is regarded as the optimal individual, indi best When there are multiple individuals with the smallest f at the same time i all Then, the individuals with higher AUC values are selected as the optimal individual Indi of the model best ；
(7) For each individual in the population Q, a binary competitive competition selection method is adopted, namely pairwise comparison is carried out, and each comparison is carried out according to the optimal individual evaluation method in the step (6) to select the individual with the smaller f i all Value of the individual into the parent population Q p ；
(8) For population Q p Each individual in the group is subjected to a crossover operation by using an analog binary crossover method. Marking the crossed population as Q x The concrete implementation process is as follows:
(8.1) generating a uniformly distributed random number β in the range of 0 to 1 for the current individual 1 When is beta 1 When the beta is less than or equal to beta, the current individual and another random individual are taken as parents to carry out subsequent operation, and when the beta is less than or equal to beta, the current individual and another random individual are taken as parents 1 >Beta, no operation is carried out;
(8.2) generating a uniformly distributed random number g in the range of 0 to 1 1 Remember k 1 The same coding position for two individuals in the cross operation;
(8.3) calculating the crossover intermediate variable β according to equation (10) g1 ；
(8.4) calculating the parent influence factor β according to equation (11) k1 ；
(8.5) Indi for two individuals 1 And Indi 2 The two filial generation individuals generated by the crossover operation are Indi c1 And Indi c2 。Indi c1 K th of (1) 1 The value of each position is calculated by the formula (12), indi c2 K of (2) 1 The values of the individual positions are calculated according to equation (13). Wherein Indi c1 (k 1 ) Is Indi c1 K of (2) 1 Value of individual position, indi 1 (k 1 ) Is Indi 1 K of (2) 1 Values of individual positions, the rest being the same;
(8.6) adding Indi c1 And Indi c2 The non-integer values are rounded to integers;
(8.7) repeating (8.1) to (8.6) steps until Q p Each individual in (a) performs a crossover operation. Putting all the obtained cross-offspring individuals into a population Q x In (1). An example of a crossover operation is given in FIG. 3, where k 1 ＝2；
(9) For population Q x Performing mutation operation by a polynomial mutation method, and marking the mutated population as Q m . The specific implementation process is as follows:
(9.1) generating a uniformly distributed random number σ in the range of 0 to 1 for the current individual 1 When σ is 1 When the sigma is less than or equal to the sigma, the subsequent operation is carried out on the individual, and when the sigma is less than or equal to the sigma 1 >When sigma is, no operation is carried out;
(9.2) generating a uniformly distributed random number g in the range of 0 to 1 2 Remember k 2 Is the coding position of the individual in the mutation operation;
(9.3) calculating the intermediate variable of variation beta according to equation (14) g2 ；
(9.4) recording k of the individual Indi 2 The upper bound of the value of the position is Indi (k) 2 ) up K of the individual Indi 2 The lower bound of the value of position is Indi (k) 2 ) low . The offspring individuals generated by mutation recording operation are Indi mut ，Indi mut K of (2) 2 The value of each position is calculated by formula (15);
Indi mut (k 2 )＝Indi(k 2 )+(Indi(k 2 ) up -Indi(k 2 ) low )×β g2 (47)
(9.5) adding Indi mut The non-integer values are rounded to integers;
(9.6) repeating (9.1) to (9.5) steps until Q x Each individual in (a) was subjected to mutation. Putting all obtained variant progeny individuals into a population Q m In (1). An example of a mutation operation is given in FIG. 4, where k 2 ＝1；
(10) The population Q p And child population Q x , child population Q m Are combined into a new population Q n ；
(11) For new population Q n And implementing an elite strategy to select and generate a next generation population. Q n Comprises an individual set Q of the previous generation population after being selected by a binary competitive bidding competition p I.e. including the best individual Indi in the previous generation population best . The elite strategy will be one after the otherOptimal individual Indi of generations best And carrying out fitness evaluation and other operations together with the new filial generation individuals so as to generate a next generation population. Recording the next generation population as Q p+1 The detailed process of the elite selection strategy is as follows:
(11.1) evaluating the new population Q according to the steps (5) and (6) n All individuals in (1) perform a fast non-dominated sorting operation to obtain Q n All non-dominating sets F of j 。
(11.2) Q p+1 As empty set, j =1;
(11.3) calculation of F j The crowdedness distance of all individuals. Note d i Is F j The crowdedness distance of the ith individual is f 1 And f 2 To F is aligned with j Will have the largest f 1 Value and maximum f 2 Two units of the value are used as boundary units, and the crowdedness d corresponding to the boundary units i Considered infinite. The crowdedness distances of the remaining individuals, excluding the boundary individuals, are calculated according to the formula (16). Wherein f is z i+1 Is shown as F j F of the (i + 1) th individual z Target fitness value, f z i-1 Is represented by F j Of (i-1) th individual z A target fitness value;
(11.4) recording of | Q p+1 L is Q p+1 Number of individuals, | F j L is F j Number of individuals in the population. When | Q p+1 |+|F j When | < N, adding F j All individuals in (1) put in Q p+1 In (1). When | Q p+1 |+|F j |>When N is present, F is j The middle individuals are sorted according to the crowding degree distance from large to small, and F after sorting j In the sequence, N- | Q is selected p+1 Putting | individual into Q p+1 Performing the following steps;
(11.5) let j = j +1;
(11.6) repeating (11.1) to (11.5) steps until | Q p+1 |＝N；
(12) Let Q = Q p+1 ，E＝E+1；
(13) Repeating the steps (5) to (12) until E = E max ；
(14) For the passage E max And (5) performing rapid non-dominated sorting and comprehensive fitness evaluation on the population Q subjected to the round evolution according to the modes of the step (5) and the step (6) to obtain the optimal individual Indi best . Combining the optimal individual Indi best And transmitting the corresponding characteristic attribution scanning block and the logistic regression model to the SAR image countercheck on-line detection module. FIG. 5 shows F obtained after the maximum evolution round 1 Pareto frontier plot of middle individuals. In this example, the optimal individual code selected is [24,117,1,3]；
(15) The SAR image confrontation sample detection data preprocessing module acquires real-time image data in the monitoring process from a target SAR system, and an online detection data set X is obtained after data normalization and data normalization processing t The online detection data set X is t Transmitting the data to an online detection module;
(16) The SAR image confrontation sample detection online detection module uses the obtained optimal characteristic attribution scanning block to calculate an online detection data set X according to the characteristic attribution expression calculation mode in the step (5) t The characteristic expression of (2); then judging X by using the obtained optimal regression model t Whether the characteristic expression of the countermeasure sample belongs to the characteristic expression of the countermeasure sample or not and sending out early warning on the found characteristic expression of the countermeasure sample; otherwise, the sample is a normal sample.
The technology of the invention and three types of existing antagonistic sample detection technologies are subjected to antagonistic sample detection experimental tests on the FUSAR-Ship data set. The experiment test selects 5 types of confrontation samples related in the technology of the invention, including FGSM, PGD, CW, deepfol and Noise; optimal individual Indi obtained by the technology of the invention best ＝[24,117,1,3](ii) a Establishing a regression model based on three types of existing countermeasure sample detection technologies, namely nuclear density and Bayesian estimation (KDBU), intrinsic dimension (LID) and mahalanobis distance fraction (MAHA), and calculating the AUC value of each type of countermeasure sample data and normal FUSAR-Ship data of the regression model; simultaneous use of the optimal feature attribution scans obtained in the present embodimentThe block processing of the challenge sample data and the normal FUSAR-Ship data involved in the experiment calculates the AUC value of the regression model obtained in the embodiment for each type of challenge sample data and normal FUSAR-Ship data, and the result is shown in FIG. 6, and we can find that: compared with the existing three traditional counterattack sample detection methods, the AUC value of the regression model obtained by the technology of the invention is in a leading position in most counterattack samples. In addition, compared with the antagonistic sample detection technology based on the Leave One Out (LOO) change of the pixel, the number of the sub-samples required by the samples with different sizes is changed as shown in FIG. 7, and it is not difficult to find: compared with the LOO technology, the technology of the invention greatly improves the calculation efficiency.
In summary, the beneficial effects of the invention are as follows: the technology of the invention can automatically obtain the optimal feature analysis granularity according to different scenes, and efficiently realize the detection of various confrontation samples in the SAR image recognition field; compared with the prior art, the SAR image confrontation detection method and device further improve the calculation efficiency and AUC performance indexes of SAR image confrontation detection.
Claims (7)
1. A synthetic aperture radar image confrontation sample detection system based on multi-target optimal characteristic attribution selection is characterized by comprising an SAR image confrontation sample detection data preprocessing module, an SAR image confrontation sample detection offline optimal characteristic attribution selection training module and an SAR image confrontation sample detection online detection module;
the SAR image confrontation sample detection data preprocessing module acquires real-time image data in a monitoring process from a real-time database of a target SAR system, obtains an online detection data set after image data normalization and image data normalization processing, and transmits the online detection data set to the SAR image confrontation sample detection online detection module; the SAR image confrontation sample detection data preprocessing module acquires historical image data from a historical database of a target SAR system, obtains standard historical image data after image data normalization and image data normalization, generates confrontation sample image data corresponding to the standard historical image data, combines the standard historical image data and the confrontation sample image data to obtain an offline training data set, and transmits the offline training data set to an SAR image confrontation sample detection offline optimal characteristic cause selection training module;
the SAR image countermeasure sample detection offline optimal feature attribution selection training module is characterized by firstly carrying out integer coding on Size, scanning interval Stride, marking whether expansion operation is carried out or not when the edge of an image is scanned and the number of model hidden Layers for feature analysis, randomly generating an initialization population, then carrying out sliding scanning on SAR image data in an offline training data set based on feature attribution scanning blocks corresponding to different individual codes through a sliding scanning submodule, carrying out feature extraction through a hidden feature extraction submodule, calculating a feature expression set corresponding to the offline training data set through a feature expression calculation submodule, carrying out logistic regression training by taking the feature expression set of the obtained offline training data set as training data through a regression model training submodule, and obtaining a logistic regression model for identifying the countermeasure sample; counting the number of sub-samples generated in the characteristic analysis process and the area under a test subject working characteristic curve of a regression model as the fitness index of an individual fitness function evaluation submodule, sequencing individuals in an initial population by adopting a quick non-dominated sequencing and fitness comprehensive evaluation submodule to obtain pareto frontier individuals, and selecting the optimal individual from the pareto frontier individuals; generating offspring populations through selection, crossing and mutation, merging the parents and the offspring to generate a new population, performing rapid non-dominated sorting and crowding distance calculation on the new population, and generating a next generation population according to the population scale; repeating the above evolution process until a maximum number of evolutionary rounds is reached, thereby obtaining individuals with pareto optima; transmitting the optimal characteristic cause scanning block and the optimal regression model corresponding to the optimal individual to an SAR image confrontation sample detection online detection module;
the SAR image confrontation sample detection online detection module performs sliding scanning on SAR image data in an online detection data set by using an optimal characteristic attribution scanning block through a sliding scanning submodule, performs characteristic extraction through an implicit characteristic extraction submodule and calculates a characteristic expression set of the online detection data set through a characteristic expression calculation submodule; then, the obtained optimal regression model is used as a countermeasure sample detection model to judge whether the feature expression of the online detection data set belongs to the feature expression of the countermeasure sample; if detecting that a certain characteristic expression belongs to the characteristic expression of the confrontation sample, sending out early warning information; otherwise, the sample is a normal sample.
2. The SAR image confrontation sample detection method based on multi-target optimal feature cause selection by applying the system of claim 1 is characterized by comprising the following steps:
(1) The SAR image confrontation sample detection data preprocessing module acquires historical SAR image data stored in a monitoring process from an SAR system as an original data set, and marks the historical SAR image data as an O; carrying out image data normalization operation on the O to obtain a normalized data set X, and then carrying out image data normalization operation through a formula (1) to obtain an offline standard data set X o ；
The SAR image data normalization module consists of segmentation operation, scaling operation and dimension reduction operation; the segmentation operation means that the original image is segmented into regions, when the input SAR historical monitoring image is a monitoring region overall image containing a plurality of identification targets, the image needs to be segmented according to the identification targets, and the segmented SAR image only contains a single identification target; when the input SAR historical image data is segmented and each SAR historical image only contains a single identification target, the segmentation operation is not needed; scaling operation means that the size of each SAR image in the data set is adjusted, and the size parameter of each SAR image is ensured to be consistent with the parameter of the target model; the dimensionality reduction operation represents that single-channel processing is carried out on the SAR image in the data set; if the input SAR historical image data is a single-channel image, the dimension reduction operation is not needed;
off-line standardData set X o Transmitting to a challenge sample generation platform for challenge sample generation, namely using 5 challenge sample generation methods for generating X o Confrontational sample data set X under target model adv (ii) a The 5 challenge sample generation methods include: FGSM attack based on fast gradient symbols, PGD attack based on gradient iteration, CW attack based on optimization, deepfol attack based on boundary decision and Noise attack based on random Noise; mixing X o And X adv And (3) segmenting according to the proportion of 4 o_train Training set X containing challenge samples adv_train Verification set X of clean data o_val Validation set X comprising challenge samples adv_val ；
(2) Setting relevant parameter values in an offline optimal feature attribution selection training module, wherein the parameter values comprise a population Size N and a maximum Size of a feature attribution scanning block max Minimum Size min Maximum step size Stride max Minimum step size Stride min Maximum number of layers selected for the feature layer L, training round EP train Cross rate beta, variance rate sigma, maximum evolution round E max ；
(3) Integer coding a set of parameters of Size, stride, panning and Layers in a feature-attributed scanning block as an individual in a multi-objective evolutionary strategy, initializing N individuals as an initial population Q, wherein the coding form of each individual is Indi = [ Size, stride, panning, layers ], indi represents any individual in the population, wherein Size represents the Size of the feature-attributed scanning block, stride represents the scanning interval of the feature-attributed scanning block, panning represents a mark of whether the feature-attributed scanning block performs an expansion operation when scanning an image edge, layers represents the number of model hidden Layers for feature analysis, and the initialization implementation processes of Size, stride, scaling and Layers are respectively shown in formulas (2) to (5):
Size＝Randint(Size min ,Size max ) (2)
Stride＝Randint(Stride min ,Stride max ) (3)
Padding＝Randint(0,1) (4)
Layers＝Randint(1,L) (5)
the random (a, b) represents that an integer with the size between a and b is randomly generated, and the value range comprises two end values of a and b; when Padding =0, the feature scanning process does not perform the edge extension operation, and when Padding =1, the feature scanning process performs the edge extension operation; mixing X o The length and the width of the medium SAR image are respectively marked as h and w, h = w under the general condition, the line scanning times of the characteristic cause scanning block is marked as m, and the calculation process of m is shown as a formula (6); the edge pixel width not covered by the line scanning process is marked as k, and the calculation process of k is shown in formula (7):
k＝(h-Size)-m×Stride (7)
wherein
Represents a round-down operation; when Padding =0, k is not equal to 0, it represents that the current Size and Stride parameter values cannot scan the function edge, and part of edge pixels of the SAR image do not participate in the subsequent feature extraction process; when Padding =1, k is not equal to 0, performing edge expansion on the SAR image before feature attribution analysis, and fixing the value of the filled pixel value to be 0, so that all pixels in the SAR image participate in the subsequent feature attribution analysis process; />
The hidden feature extraction operation can be used for sequentially extracting the output features of the hidden Layers from the last hidden layer of the target model according to the number of Layers until the number of the extracted hidden Layers is equal to the number of Layers;
(4) Marking the current evolution turn of the multi-objective optimization technology as E, and enabling E =0;
(5) Evaluating fitness function of initial population QNamely, the SAR image confrontation sample detection offline optimal characteristic attribution selection training module firstly respectively calculates a data set X according to the characteristic attribution scanning blocks corresponding to the individuals in Q o_train ，X adv_train ，X o_val ，X adv_val The characteristic expression of (2); then, using training set X o_train ，X adv_train Performing logistic regression training on the feature expression set, wherein the training turn is EP train Using verification set X in the training process o_val ，X adv_val The obtained regression model is verified by the characteristic expression set, the AUC value of the regression model is calculated, and the number of the sub-samples generated in the characteristic analysis process is evaluated;
(6) Carrying out rapid non-dominated sorting and comprehensive fitness evaluation on individuals in the population Q; the fast non-dominated sorting is to search the pareto frontier of the population Q by measuring the advantages and disadvantages of the individuals in two evaluation indexes of an AUC value and the number of sub samples, and the fitness comprehensively evaluates the two evaluation indexes in the pareto frontier individual set to select the optimal individual of the population Q; in specific operation, the negative number and the sub-sample number of the AUC value of an individual are taken as optimization objective functions and respectively marked as f 1 And f 2 (ii) a Note n i Number of individuals dominating the ith individual in the population, S i A set of individuals dominated by the ith individual; what an individual dominates over another is: for the AUC value corresponding to the individual and the number of subsamples, the dominant individual performs better than the dominated individual, namely, the model AUC value is higher than the dominated individual, and the number of subsamples is smaller than the dominated individual;
(7) For each individual in the population Q, a binary competitive competition selection method is adopted, namely pairwise comparison is carried out, and the optimal individual evaluation method in the step (6) is used for selecting individuals with smaller f in each comparison i all The individuals with the value enter the parent population Q p ；
(8) For population Q p Each individual in the system is subjected to cross operation by adopting a simulated binary cross method; marking the crossed population as Q x ；
(9) For population Q x Is varied by a polynomial variation methodPerforming different operations, and marking the varied population as Q m ；
(10) The population Q p And child population Q x , child population Q m Are combined into a new population Q n ；
(11) For new population Q n Implementing an elite strategy to select and generate a next generation population; q n Comprises an individual set Q of a previous generation population after being selected by a binary competitive bidding competition p I.e. including the best individual Indi in the previous generation population best (ii) a Elite strategy will optimize the previous generation of individuals, indi best Carrying out fitness evaluation and other operations with the new filial generation individuals to generate a next generation population; recording the next generation population as Q p+1 ；
(12) Let Q = Q p+1 ，E＝E+1；
(13) Repeating the steps (5) to (12) until E = E max ；
(14) For the passage E max And (5) performing rapid non-dominated sorting and comprehensive fitness evaluation on the population Q subjected to the round evolution according to the modes of the step (5) and the step (6) to obtain the optimal individual Indi best (ii) a Combining the optimal individual Indi best The corresponding characteristic cause scanning block and the logistic regression model are transmitted to an SAR image confrontation sample detection online detection module;
(15) The SAR image confrontation sample detection data preprocessing module acquires real-time image data in the monitoring process from a target SAR system, and an online detection data set X is obtained after data normalization and data normalization processing t The online detection data set X is t Transmitting the data to an online detection module;
(16) The SAR image confrontation sample detection online detection module uses the obtained optimal characteristic attribution scanning block to calculate an online detection data set X according to the characteristic attribution expression calculation mode in the step (5) t The characteristic expression of (2); then judging X by using the obtained optimal regression model t Whether the feature expression of the confrontation sample belongs to the feature expression of the confrontation sample or not, and sending out early warning on the found feature expression of the confrontation sample; otherwise, the sample is a normal sample.
3. The SAR image confrontation sample detection method according to claim 2, characterized in that the specific calculation process of the step (5) is as follows:
(5.1) let i =0, label the maximum number of images of the dataset as i max ；
(5.2) setting a set F of subsample images sub Selecting the ith image x of the data set as a current feature analysis image for an empty set, and performing sliding scanning on the image by using a feature attribution scanning block; starting from the (0, 0) position in the x image pixel matrix, the scanning operation selects a square area with the Size of Size in the x image pixel matrix according to the Size parameter to perform sub-sample generation operation; the sub-sample generation operation clones a new image with the same value as the original x pixel value, sets the pixel value of the scanning area of the new image to 0, and then adds the image into the sub-sample image set F sub Performing the following steps; after one-time sub-sample generation operation is completed, the characteristic attribution scanning block slides in rows according to the Stride parameters, and one-time sub-sample generation operation is performed in each sliding scanning; repeating the process until the next scanning of the scanning block exceeds the width of the pixel matrix, moving the scanning block to the next row scanning position in the x image pixel matrix according to the Stride parameters, and starting sliding scanning according to rows again; repeating the above process until the feature cause scanning block completes all the line-by-line scanning operations in the x matrix pixels; finally, the original image x is also added into the sub-sample image set F sub Performing the following steps; the starting point position of the p-th scanning of the l-th line in the sliding scanning operation can be recorded as (l × Stride, p × Stride), and the coordinates of the four points of the corresponding square scanning area are respectively: (l × Stride, p × Stride), (l × Stride, p × Stride + Size), (l × Stride + Size, p × Stride + Size);
(5.3) set F of subsamples of the image sub Inputting a target model, selecting a corresponding model hidden layer according to a given Layers parameter, and obtaining the output of each image in the subsample set corresponding to the hidden layer in the model; carrying out average pooling operation on the output of the hidden layer to obtain a subsample set F sub A corresponding set of implicit features;
(5.4) subtracting the implicit characteristic corresponding to the original image x from the implicit characteristic corresponding to the sub-sample image except the original image x in the implicit characteristic set to obtain an image characteristic change matrix capable of representing the influence of image pixel change;
(5.5) carrying out quartile distance (IQR) statistics on the image characteristic change matrix to obtain a characteristic expression vector capable of representing the image;
(5.6) let i = i +1;
(5.7) repeating (5.2) to (5.6) steps until i = i max ；
(5.8) merging all the feature expression vectors obtained in the step (5.7) to obtain a feature expression set corresponding to the data set;
the specific implementation process of the AUC value evaluation is as follows:
verification set X o_val The number of samples in (1) is R, and the samples are marked as positive samples; verification set X adv_val The number of samples in (1) is T, and the samples are marked as negative samples; the prediction score of the verification set sample in the regression model is marked as P, P Positive sample Representing the prediction score, P, of the regression model on a single positive sample Negative sample Representing the predicted score of the regression model for a single negative sample; note I (P) Positive sample ,P Negative sample ) Predicting an evaluation value for a sample of a positive and negative sample pair; when P is present Positive sample >P Negative sample When is, I (P) Positive sample ,P Negative sample ) =1; when P is present Positive sample ＝P Negative sample When is, I (P) Positive sample ,P Negative sample ) =0.5; when P is present Positive sample <P Negative sample When is, I (P) Positive sample ,P Negative sample ) =0; calculating the sample prediction evaluation value of all the positive and negative sample pairs, and calculating the AUC value of the regression model to the verification set sample according to the formula (8):
the specific implementation process of calculating the number of the sub-samples generated in the feature analysis process is as follows: attributing scan block parameters Size, stride, padding using features corresponding to individuals, and calculating k and m corresponding to the current individual through formula (6) and formula (7); when Padding =0, the number of individually corresponding subsamples is m × m +1; when Padding =1, if k =0, the individually corresponding number of subsamples is m × m +1; when Padding =1, if k ≠ 0, the number of individually corresponding subsamples is (m + 1) × (m + 1) +1.
4. The SAR image confrontation sample detection method according to claim 2, wherein the step (6) is realized by the following specific steps:
(6.1) traversing all individuals of the current population Q, and calculating n corresponding to each individual i Value and set S of individuals dominated by the individual i ；
(6.2) mixing all n i The individuals of =0 are stored in the set F 1 Performing the following steps;
(6.3) let j =1;
(6.4) letting H be an empty set;
(6.5) traverse F j All individuals in (1), note S u Is as a quilt F j The set of individuals governed by the u-th individual, n q To dominate S u Number of subjects in qth; to obtain F j S corresponding to each individual in u ；
(6.6) for all S u Go through each S u Calculating n corresponding to each individual q A value of and let n q ＝n q -1；
(6.7) if n is q If not less than 0, then S u The q individuals in (1) are placed in a set H;
(6.8) let j = j +1;
(6.9) order F j ＝H；
(6.10) repeating (6.4) to (6.9) steps until F is obtained j Is an empty set;
(6.11)F 1 all of them are dominant ones, and F 1 The individual in (1) is used as a pareto frontier individual of a population Q, the optimal individual is selected by integrating two evaluation indexes, and the specific selection process is as follows: when F is present 1 Only contains 1 individual, the individual is selected as the optimal individual Indi in the evaluation best (ii) a When F is present 1 2 individuals, the individual with the higher AUC value is selected as the optimal individual Indi in the evaluation best (ii) a When F is present 1 When 3 or more than 3 individuals are included in the evaluation data, the fitness comprehensive evaluation value f of the ith individual is calculated according to the formula (9) i all Wherein f is 1 i F representing the ith individual 1 Optimization of the target value, f 2 i F representing the ith individual 2 Optimizing the target value; selection f i all The smallest individual is regarded as the optimal individual, indi best When there are multiple individuals with the smallest f at the same time i all When the individuals with higher AUC values are selected as the optimal individual Indi of the model best ；
5. The SAR image confrontation sample detection method according to claim 2, wherein the step (8) is realized by the following steps:
(8.1) generating a uniformly distributed random number β in the range of 0 to 1 for the current individual 1 When is beta 1 When the beta is less than or equal to beta, the current individual and another random individual are taken as parents to carry out subsequent operation, and when the beta is less than or equal to beta, the current individual and another random individual are taken as parents 1 >Beta, no operation is carried out;
(8.2) generating a uniformly distributed random number g in the range of 0 to 1 1 Remember k 1 The same coding position for two individuals in the cross operation;
(8.3) calculating the crossover intermediate variable β according to equation (10) g1 ；
(8.4) calculating the parent influence factor β according to equation (11) k1 ；
(8.5) Indi for two individuals 1 And Indi 2 The two filial generation individuals generated by the crossover operation are Indi c1 And Indi c2 ；Indi c1 K of (2) 1 The value of each position is calculated by the formula (12), indi c2 K of (2) 1 The value of each position is calculated according to the formula (13); wherein Indi c1 (k 1 ) Is Indi c1 K of (2) 1 Value of individual position, indi 1 (k 1 ) Is Indi 1 K of (2) 1 Values of individual positions, the rest being the same;
(8.6) adding Indi c1 And Indi c2 The non-integer values are rounded to integers;
(8.7) repeating the steps (8.1) to (8.6) until Q p Each individual in (a) has performed a crossover operation; putting all the obtained cross-offspring individuals into a population Q x In (1).
6. The SAR image confrontation sample detection method according to claim 2, wherein the step (9) is implemented by the following steps:
(9.1) generating a uniformly distributed random number σ in the range of 0 to 1 for the current individual 1 When σ is 1 When the sigma is less than or equal to the sigma, the subsequent operation is carried out on the individual, and when the sigma is less than or equal to the sigma 1 >When sigma is, no operation is carried out;
(9.2) generating a uniformly distributed random number g in the range of 0 to 1 2 Remember k 2 The coding position of an individual in the mutation operation;
(9.3) calculating the intermediate variable of variation beta according to equation (14) g2 ；
(9.4) recording k of the individual Indi 2 The upper bound of the value of the position is Indi (k) 2 ) up K of Individual Indi 2 The lower bound of the value of position is Indi (k) 2 ) low (ii) a The offspring generated by the mutation recording operation is Indi mut ，Indi mut K th of (1) 2 The value of each position is calculated by formula (15);
Indi mut (k 2 )＝Indi(k 2 )+(Indi(k 2 ) up -Indi(k 2 ) low )×β g2 (15)
(9.5) adding Indi mut The non-integer values are rounded to integers;
(9.6) repeating (9.1) to (9.5) steps until Q x Each individual in (a) has undergone a mutation operation; putting all obtained variant progeny individuals into a population Q m In (1).
7. The SAR image confrontation sample detection method according to claim 2, characterized in that the detailed process of the elite selection strategy is as follows:
(11.1) evaluating the new population Q according to the steps (5) and (6) n All individuals in (1) perform a fast non-dominated sorting operation to obtain Q n All non-dominant sets F of j ；
(11.2) Q p+1 As empty set, j =1;
(11.3) calculation of F j The crowdedness distance of all individuals; note d i Is F j The congestion degree distance of the ith individual is defined as f 1 And f 2 To F j Will have the largest f 1 Value and maximum f 2 Two units of the value are used as boundary units, and the crowdedness d corresponding to the boundary units i Considered infinite;calculating the crowding degree distances of other individuals except the boundary individuals according to a formula (16); wherein f is z i+1 Is represented by F j Of (i + 1) th individual z Target fitness value, f z i-1 Is shown as F j Of (i-1) th individual z A target fitness value;
(11.4) recording of | Q p+1 L is Q p+1 Number of individuals, | F j L is F j The number of individuals; when | Q p+1 |+|F j When | < N, adding F j All individuals in (1) put in Q p+1 Performing the following steps; when | Q p+1 |+|F j |>When N is present, F is j The middle individuals are sorted from large to small according to the crowdedness distance, and F is arranged after sorting j In the sequence, N- | Q is selected p+1 | Individual put in Q p+1 Performing the following steps;
(11.5) let j = j +1;
(11.6) repeating the steps (11.1) to (11.5) until the value of | Q p+1 |＝N。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211504977.8A CN115861804A (en) | 2022-11-28 | 2022-11-28 | SAR image countercheck sample detection system and method based on optimal feature attribution selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211504977.8A CN115861804A (en) | 2022-11-28 | 2022-11-28 | SAR image countercheck sample detection system and method based on optimal feature attribution selection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115861804A true CN115861804A (en) | 2023-03-28 |
Family
ID=85667378
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211504977.8A Pending CN115861804A (en) | 2022-11-28 | 2022-11-28 | SAR image countercheck sample detection system and method based on optimal feature attribution selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115861804A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116882293A (en) * | 2023-07-18 | 2023-10-13 | 广东明阳电气股份有限公司 | Multi-objective optimization method and device for transformer and storage medium |
-
2022
- 2022-11-28 CN CN202211504977.8A patent/CN115861804A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116882293A (en) * | 2023-07-18 | 2023-10-13 | 广东明阳电气股份有限公司 | Multi-objective optimization method and device for transformer and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bendale et al. | Towards open set deep networks | |
| CN111814902A (en) | Target detection model training method, target identification method, device and medium | |
| CN111126134A (en) | Radar radiation source deep learning identification method based on non-fingerprint signal eliminator | |
| JP2008262331A (en) | Object tracking device and object tracking method | |
| CN114694178A (en) | Method and system for monitoring safety helmet in power operation based on fast-RCNN algorithm | |
| Mitrakis et al. | Decision fusion of GA self-organizing neuro-fuzzy multilayered classifiers for land cover classification using textural and spectral features | |
| CN115861804A (en) | SAR image countercheck sample detection system and method based on optimal feature attribution selection | |
| CN107766792B (en) | Remote sensing image ship target identification method | |
| Li et al. | SEENN: Towards Temporal Spiking Early Exit Neural Networks | |
| CN116580322A (en) | Unmanned aerial vehicle infrared small target detection method under ground background | |
| CN111144462A (en) | Unknown individual identification method and device for radar signals | |
| CN115079116A (en) | Radar target identification method based on Transformer and time convolution network | |
| CN114037001A (en) | Mechanical pump small sample fault diagnosis method based on WGAN-GP-C and metric learning | |
| CN113537240B (en) | Deformation zone intelligent extraction method and system based on radar sequence image | |
| CN115761888A (en) | Tower crane operator abnormal behavior detection method based on NL-C3D model | |
| CN115359248A (en) | Robot navigation obstacle avoidance method and system based on meta-learning | |
| CN115310559A (en) | Target identification method based on Bayes-Transformer neural network model | |
| Patsei et al. | Multi-class object classification model based on error-correcting output codes | |
| CN114202671A (en) | Image prediction optimization processing method and device | |
| CN113269217A (en) | Radar target classification method based on Fisher criterion | |
| CN117390688B (en) | Model inversion method based on supervision training | |
| EP4357977A1 (en) | Device and method for determining an encoder configured image analysis | |
| CN117808650B (en) | Precipitation prediction method based on Transform-Flownet and R-FPN | |
| Pandya et al. | A novel approach for vehicle detection and classification | |
| CN116109926A (en) | Robust SAR image recognition system and method integrating feature loss and three-objective optimization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |